Apparatus and method for recognizing speech based on a deep-neural-network (DNN) sound model

Abstract

A speech recognition apparatus based on a deep-neural-network (DNN) sound model includes a memory and a processor. As the processor executes a program stored in the memory, the processor generates sound-model state sets corresponding to a plurality of pieces of set training speech data included in multi-set training speech data, generates a multi-set state cluster from the sound-model state sets, and sets the multi-set training speech data as an input node and the multi-set state cluster as output nodes so as to learn a DNN structured parameter.

Description

CROSS-REFERENCE TO RELATED APPLICATION[0001]This application claims priority to and the benefit of Korean Patent Application No. 10-2006-0005755, filed on Jan. 18, 2016, the disclosure of which is incorporated herein by reference in its entirety.BACKGROUND[0002]1. Field of the Invention[0003]The present invention relates to an apparatus and method for recognizing speech, and more particularly, to a speech recognition apparatus and method based on a deep-neural-network (DNN) sound model.[0004]2. Discussion of Related Art[0005]A context-dependent deep-neural-network (DNN)-hidden-Markov-Model (HMM) technique using a combination of a DNN and an HMM has been actively applied to sound models for speech recognition by replacing an existing CD-Gaussian-mixture-model-HMM (CD-GMM-HMM) (hereinafter referred to as ‘GMM-HMM’) technique.[0006]A DNN-HMM technique according to the related art is performed as will be described below.[0007]First, a state of an HMM corresponding to an output node or a...

AI Technical Summary

Benefits of technology

The present invention provides a method for learning a sound model for speech recognition using multi-set training speech data with different characteristics. The method involves generating sound-model state sets corresponding to the training speech data, setting a multi-set state cluster to learn a deep-neural-network (DNN) structural parameter, and using the learned DNN structural parameter to recognize user's speech through a DNN-Hidden-Markov-model (HMM) structure. Additionally, the invention provides a speech recognition apparatus and method for performing speech recognition by setting a sound-model state set corresponding to characteristic information of a user's speech. The technical effects of the invention include improved accuracy and efficiency in speech recognition using multi-set training speech data with different characteristics.

Problems solved by technology

However, in a DNN learning technique according to the related art, state-level alignment information is determined beforehand and thus an output node of a DNN structure cannot be changed.

However, it is inefficient to determine a state of large-size training speech data having different acoustic-statistical characteristics (e.g., a sound model for recognizing English speech of multiple native speakers who can speak, for example, Chinese, Korean, and English) using one decision tree.

However, a DNN-HMM-based learning technique according to the related art is a method of learning a structure having characteristics and parameters that most appropriately discriminate predetermined states and thus is not applicable in the field of speech recognition application performed with respect to multiple native speakers having different acoustic-statistical characteristics.

Images